Abstract:
Cyber-physical systems (CPS), such as robots and self-driving cars, demand rigorous scheduling to prevent failure, where every millisecond of processing time can be critical. These systems often rely on heterogeneous computing environments, which include CPUs, GPUs, and specialized accelerators, to meet their computational needs efficiently. However, leveraging these diverse processing units to fulfill strict physical constraints remains a significant challenge, as existing scheduling solutions often fall short of addressing the complexities involved in a comprehensive manner.
This talk by Dr. Mehmet E. Belviranli delves into the intricate world of creating efficient compute schedules for CPS that not only cater to their diverse computational hardware but also adhere to real-world constraints critical for system safety. We begin by examining the role of neural network (NN) inference in CPS, exploring strategies to balance energy consumption, latency and throughput by distributing the layers of NN across different accelerators. We then introduce a novel, end-to-end framework that integrates physical constraints, heterogeneous computational resources, and latency considerations into a cohesive mixed-integer linear problem, demonstrating through case studies how this approach yields optimal scheduling solutions under varied conditions.
Through this exploration, we aim to shed light on the untapped potential of heterogeneous computing in enhancing the reliability and performance of CPS. We will also outline future directions in developing a more robust ecosystem for these complex computing environments, highlighting our contribution to this evolving field.
Speaker(s): Dr. Belviranli
Virtual: https://events.vtools.ieee.org/m/417666